An Interesting Development - New CMP architecture

I have zero interest in mining, but will be interested to see whether CMP means removal of Tensor/FP units in order to add more integer/logic functionality.

EDIT: Upon reflection, maybe CMP will just be RTX3XXX chips binned for broken video outputs. Comparing an unofficial early RTX3060 result:

with official CMP specs: A Crypto Mining GPU for Professionals | NVIDIA doesn’t look too inspiring…

RTX 3060 software drivers are designed to detect specific attributes of the Ethereum cryptocurrency mining algorithm, and limit the hash rate, or cryptocurrency mining efficiency, by around 50 percent.

I fully anticipate the law of unintended consequences to kick in, with users bitterly complaining in these forums about their code (not mining!) underperforming on the RTX 3060. Cyrptocurrency mining appears to bring out the worst in people and corporations.

So is the CMP only useful for mining, or could it be used like a traditional non-display Tesla card for raw CUDA / AI / ML programming? Something like an Ampere continuation of the T4? Not really inclined to get warm with the A100,

Impossible to say at the current stage. NVIDIA wouldn’t want to cannibalize RTX 3060 sales by allowing the CMP chips to be useful for too many other applications besides mining.

NVIDIA’s announcement makes it sound like the CMP chips are dies that didn’t qualify for use in the RTX 3060, e.g. have defects in the video-out part or or malfunction at the RTX 3060 target frequency. Anybody’s guess as to what else might be non-functional on these chips. DP units perhaps, some of the caches?

It sounds like the RTX 3060 “crypto brake” is going to use application fingerprinting either statically by kernel code analysis or dynamically by performance counter data. Anybody’s guess as to how many false positives this will result in, e.g. when running other, non-mining, crypto applications. Whatever the finger-printing does, I would expect workarounds to pop up, not by hacking NVIDIA’s software but by changing the characteristics of crypto-mining software (“cloaking”) to evade the checks at a minimal cost in performance.

“NVIDIA wouldn’t want to cannibalize RTX 3060 sales” – I think they are. Given the supply shortages, there’s more involved than just the chips. With shortages in GDDR6 memory you are cannibalizing on 3060 sales.

Yes, I also think there will be false positives on the 3060, and people will be able to circumvent the throttling (hack the BIOS, although it’s apparently encrypted, use old drivers, etc.). But if that takes half a year, it will have become useless for mining purposes. Even an unthrottled 3060 will be useless for mining in half a year.
Not my concern, not interested in getting a 3060 or mining, more interested in getting alternatives to A100 (price), T4 (performance), RTX Titan (price – still 2.5k), RTX 4000 (performance). And your answer indicates that the CMP won’t be useful for anything other than mining. I find it ironic that I’m back to the 2080 Ti for AI / ML purposes. They’re now a grand on ebay and obtainable – unlike the 3080s and 3090s.

I speculated that this might be the case. I also stated that it is too early to tell for sure, as the exact specifications of these products are not known yet. By “cannibalize” I meant pre-empting sales of RTX 3060s that are actually available for sale, not ones NVIDIA can’t make due to production restrictions. This is a fine line to tread for NVIDIA, because their goal is presumably to reduce demand for the RTX 3060 to pretty much exactly the level at which they can produce the card, but not significantly more.

Some slight official elaboration here, "According to Mr. Del Rizzo: “It’s not just a driver thing. There is a secure handshake between the driver, the RTX 3060 silicon, and the BIOS (firmware) that prevents removal of the hash rate limiter.”

All of the following is speculation: The whole mechanism presumably comes in two parts: mining detection and rate limiting. The rate limiting is presumably a function of the VBIOS and the hardware. As far as I know, crypto-mining involves a lot of bit-twiddling, so there might be an internal switch that limits the rate at which LOP3s or shifts can be issued. The driver portion presumably has the task of detecting crypto mining. This part needs to be easy to update to minimize false positive and false negative detections as the mining software evolves and will be continuously tuned between driver release, possibly even with the help of some AI mechanism.

To continue to flog a dead horse of mine from other threads, I wish Nvidia would do the job properly and design an architecture for the task.

I don’t believe crypto, as with any number of GPGPU tasks, makes significant use of FP or Tensor units. If this representation of one quarter of a GA10X SM is to scale, removal of this functionality would result in significant room for integer/logic expansion.

CMP would seem to be an expedient way to utilize non RTX3XXX qualifying hardware to address a potentially transient mining demand.

One can live in hope however…


I would be highly surprised if that were the case.

Well, looking at the Nvidia GA102 Whitepaper, page 19, the Tensor and RT cores are taking a significant amount of real estate.

EDIT: Perhaps some artistic license involved there.

The die picture on page 19 of that document is that of a Turing class chip (likely TU102), and the floor plan partitioning suggested by the color-shaded areas is purely symbolic. So yes, artistic license.

From yesterdays earnings transcript: the CEO says: “And so we expect it to be a small part of our business as we go forward”, and they expect sales to be a mere 50 million.
It causes me to believe that if the CMP cards could be used for more than just mining, they have a good chance to not be trashware in the dumpster in half a year and could perhaps even have some resale value. In a way, we already know they don’t have a long span. Even though they can be used for more than just Ether (but so can the 3060), if we can’t do more with them than just mining, we’ll just make Greta happy as they’re basically just throw-away-stuff at that point.
Also, I read in the financial news that Taiwan Semiconductors (a foundry NVidia uses) apparently only works with AMD going forward. That will also have an impact on price and availability.

As I see it, the main environmental issue with cryptomining using CMP cards is not the disposal of these card after a relatively short amount of time (“trashware”). It’s the energy needed to operate cryptomining farms that use these cards. As far as the GPU chips themselves are concerned, those would have gone straight to recycling or landfill if they weren’t salvaged for CMP cards. Now they are taking a detour through a cryptomining farm. Arguably this is “re-use”.

Re “they expect sales to be a mere 50 million”: Does that refer to 50 million units or a sales volume of 50 million dollars? Regardless of which it is, I don’t buy the story that the introduction of the CMP products is going to “help gamers”. Occam’s Razor suggests it’s about converting chips that would otherwise be trash into a sellable product that drives incremental revenue.

50 million $. Sales numbers are always in currency, unless indicated otherwise.

Not disagreeing with your points. But I think it’s smart to “re-purpose” stuff that otherwise would be useless. If that increases sales/earnings/revenue/profit for NVidia, I have nothing against that (I have a lot of NVDA shares). I wish they did that for “unusable” chips for the 3090 and, in the future, Lovelace chips, so we have better non-display cards for AI / ML / NNs – basically for some Tesla continuation. I still find the A100 too pricey.

Another data point on the CMP architectures.

From the TechPowerUp post:

The 50HX and 90HX cards are expected to launch later this year and will likely restrict GPU supply if the cryptocurrency boom continues.

I do not understand where those concerns are coming from, under the reasonable assumption that all CMP chips are chips that would otherwise have gone straight to recycling / trash because they are either (1) too slow (2) have certain functional defects. It would be illogical from a revenue perspective for NVIDIA to sell chips as cheap CMPs that could have otherwise been sold as incrementally more expensive GTXs.

I believe there is potential risk that CMP chips won’t help alleviate supply constraints in the gaming GPU market because CMP chips have sufficiently unattractive price / performance characteristics compared to popular gaming GPUs, so cyrptominers will want to continue using GTX cards. But I fail to see how the introduction of CMP chips would exacerbate shortages in the GPU market.

Anyone suffering the above will now be able to search for driver 470.05 in order to confirm whether it is indeed caused by the throttling code or not:


Plot twist!

FWIW, that thread in the TechPowerUp forums has decent short-time entertainment value.